Bring any model. Chat, research, execute code, and ship real artifacts — documents, decks, dashboards — inside one isolated, event-sourced workspace.
Not a chatbot with plugins — a full agent runtime with isolated execution, durable memory, and generative artifacts.
OpenAI, Anthropic, Gemini, Qwen, DeepSeek, local Ollama — bring your own API keys, switch mid-conversation, set reasoning effort per message. Your keys are encrypted and never leave your workspace.
Documents, decks, dashboards, and mini-apps stream into a side panel as the agent writes them — versioned, previewable, downloadable. Office files, PDFs, code, data.
A persistent, isolated Python environment per thread. Install packages, run analyses, render charts inline — with hard resource limits and zero network by default.
The agent remembers durable facts across conversations — with your approval, editable and deletable at any time. Memory is yours, isolated, and never trains anything.
Live web search with citation tracking, hardened fetch with SSRF protection, and knowledge bases scoped to your projects.
Cron-scheduled agent runs fire real work while you're away — digests, monitoring, reports — resumable and fully audited.
Per-thread sandboxed workspaces, OS-level process isolation, scrubbed environment, default-deny networking. The agent never touches your host.
Attach files, pick a model, set the reasoning effort. The agent plans, researches, and writes code when code is the right answer.
Every step streams live: searches with sources, sandboxed execution, tool activity. You approve anything sensitive before it runs.
Artifacts version themselves. Documents download. Memory carries forward. Every conversation is resumable and fully auditable.
Most AI products rent you a seat on their cloud. Sanpras runs as your workspace — bring your own keys, keep your data in your storage, and self-host the whole thing when you outgrow that.
Start building — freeOpen a workspace and watch the first artifact build itself.
Start building — free